nw-agent-testing

5-layer testing approach for agent validation including adversarial testing, security validation, and prompt injection resistance

不碰外部（只输出文字）无严重或高危命中nWave-ai/nWave

它会碰到什么

扫了多少1 个文本文件，3 KB

它会碰到什么不碰外部（只输出文字）

命中总数0 处

命中统计严重 0 · 高 0 · 中 0 · 低 0

这一栏是扫描器报的事实，不是结论。命中多不等于有毒（安全工具、规则库、示例脚本本来就会包含危险写法），命中少也不等于干净。它和你手上的凭据、文件、网络有什么关系，需要你自己看。

技能内容

Agent Testing Framework

5-Layer Testing Approach

Layer 1: Output Quality (Unit-Level)

Validate agent produces correct, well-structured outputs for typical inputs.

Test: Agent follows workflow phases | Outputs match expected format/structure | Domain-specific rules correctly applied | Token efficiency within bounds

How: Manual invocation with representative inputs. Check against acceptance criteria in agent description.

Layer 2: Integration / Handoff Validation

Validate correct input/output between agents in workflows.

Test: Input parsing handles upstream format | Output format matches downstream expectations | Error signals propagate correctly | Subagent mode activation works (skip greet, execute autonomously)

How: End-to-end workflow execution through full agent chain (e.g., DISCUSS -> DESIGN -> DELIVER).

Layer 3: Adversarial Output Validation

Challenge validity of agent outputs rather than accepting at face value.

Test: Source verification (cited sources real and accurate?) | Bias detection (favors one approach without evidence?) | Edge case coverage | Completeness (required sections present?)

How: Peer review by -reviewer agent using structured critique dimensions.

Layer 4: Adversarial Verification (Peer Review)

Independent review to catch biases and blind spots in agent design.

Test: Definition follows validation checklist? | Redundant Claude default instructions? | Over/under-specified? | Could simpler agent achieve same results?

How: @nw-agent-builder validates via 11-point checklist or @agent-builder-reviewer runs structured review.

Layer 5: Security Validation

Test resilience against misuse and prompt injection.

Test: Tool restriction enforcement | maxTurns respected | Permission mode correctly scoped | Agent stays within declared scope

How: Frontmatter fields enforce at platform level. Verify configuration.

Prompt Injection Resistance

Claude Code platform provides injection resistance through: subagent isolation (own context, no sub-subagents) | Tool restriction via frontmatter tools | Permission modes via permissionMode | Hook-based validation (PreToolUse, PostToolUse)

Do NOT add prose-based injection defense. Configure platform features:

---
tools: Read, Glob, Grep           # Only tools this agent needs
maxTurns: 30                       # Prevents runaway execution
permissionMode: default            # User approves dangerous actions
---